Deep learning is currently playing an essential role toward intelligent fault diagnosis. Nevertheless, the automatically learned representations often suffer from a lack of interpretability. This paper proposes a denoising fused wavelets net (DFWNet) for aeroengine bevel gear fault diagnosis with improved model performance and interpretability. In contrast to standard convolutional neural network, the convolutional kernel is replaced by wavelet basis, and only scale parameters of the wavelet are directly learned from vibration data in wavelet convolution. To enhance the feature learning ability and alleviate the noise impact, learnable thresholds are used for soft thresholding denoising and weights based on energy-to-entropy ratio are given to each channel. Experiment study conducted on an aeroengine bevel gear fault dataset proves that the proposed approach converges faster and performs better with interpretable kernels.